Path Diversity for Enhanced Media Streaming
John G. Apostolopoulos, Mitchell D. Trott
Mobile and Media Systems Laboratory
HP Laboratories Palo Alto
HPL-2004-123
July 14, 2004*
video streaming,
path diversity,
multipath
streaming, multiple
description coding
Media streaming over best-effort packet networks such as the Internet is
quite challenging because of the dynamic and unpredictable available
bandwidth, loss rate, and delay. Recently, streaming over multiple paths
to provide path diversity has emerged as an approach to help overcome
these problems. This paper provides an overview of the benefits and use
of path diversity for media streaming. The different approaches for media
coding and streaming over multiple paths are examined, together with
architectures for achieving path diversity between single or multiple
senders and a single receiver. Important examples include using the
distributed servers in a content delivery network to provide path diversity
to a requesting client, using multiple 802.11 wireless access points to
provide path diversity to a mobile client, and using relays to provide lowlatency media communication. The design, analysis, and operation of
media streaming systems that use path diversity are considered, with
emphasis on the accurate performance models needed to select the best
paths or best servers.
* Internal Accession Date Only
Approved for External Publication
 Copyright IEEE. To be published in IEEE Communications Magazine, special issue on Proxy Support for
Streaming on the Internet, August 2004.
IEEE Communications Magazine, special issue on “Proxy Support for Streaming on the Internet”, August 2004.
PATH DIVERSITY FOR ENHANCED MEDIA STREAMING
John G. Apostolopoulos and Mitchell D. Trott
Streaming Media Systems Group
Hewlett-Packard Labs, Palo Alto, CA 94304
ABSTRACT
Media streaming over best-effort packet networks such as
the Internet is quite challenging because of the dynamic and
unpredictable available bandwidth, loss rate, and delay. Recently, streaming over multiple paths to provide path diversity has emerged as an approach to help overcome these
problems. This paper provides an overview of the benefits and use of path diversity for media streaming. The
different approaches for media coding and streaming over
multiple paths are examined, together with architectures for
achieving path diversity between single or multiple senders
and a single receiver. Important examples include using the
distributed servers in a content delivery network to provide
path diversity to a requesting client, using multiple 802.11
wireless access points to provide path diversity to a mobile
client, and using relays to provide low-latency media communication. The design, analysis, and operation of media
streaming systems that use path diversity are considered,
with emphasis on the accurate performance models needed
to select the best paths or best servers.
1. INTRODUCTION
Path diversity is a transmission technique that sends data
simultaneously through two or more paths in a packet-based
network. The paths may originate from single [1, 2, 3, 4, 5,
6] or multiple sources [7, 8, 9]. The single-source case in a
typical video streaming application is shown in Figure 1 [1].
A more complex multiple-source scenario is illustrated in
Figure 2 [8].
Original
Video
Path # 1
MD
#1
Relay 1
MD #1
MD Video
Encoder
MD #2
Path
Selector
Packet
Receiver
MD
#2
MD Video
Decoder
Path # 2
Relay 2
MD Video Encoder
and Transmitter
Received
Packets
Packet Network
and Relay Infrastructure
Reconstructed
Video
MD Video Receiver
and Decoder
Fig. 1. Path diversity: media packets are sent over multiple
paths in a packet network. In this case different media descriptions are explicitly sent over different paths via relays.
C
Server
Server
Server
C CC
C
C
C C
C
C
Server
Server
Server
CC
C
C
C = Client
Fig. 2. Path diversity in a content delivery network: multiple servers stream content to each client. A basic question:
one green and one blue server should be selected to stream
to the clients—which blue server should be selected? The
closer one which shares a link with the green server, or the
further blue server?
The use of multiple paths through the transport network
for streaming has been proposed to help overcome the loss [1]
and delay [2] problems that afflict streaming media and lowlatency communication. In addition, it has long been known
that multiple paths can improve fault tolerance and link recovery for data delivery, as well as provide larger aggregate
bandwidth, load balancing, and faster bulk data downloads.
Diversity techniques have been studied for many years
for wireless communication, e.g., frequency, time, and spatial diversity. However, path diversity over packet networks
has received limited attention until recently. The early work
in this area, e.g. [10, 11], focuses on reliable data delivery through error correction coding or retransmission, and
is generally performed in the context of multiple virtual circuits on connection-oriented systems such as ATM.
A number of thorough experimental studies have demonstrated the great variability in the end-to-end performance
observed over the Internet. This variability is analogous to
the variability of wireless links, and motivates the application of wireless diversity techniques to wired and wireless
Internet communication. Further motivation for the use of
multiple paths is that the default (single) path between two
nodes on a network is often not the best path. For example,
the measurement study [12] comparing the paths between
two hosts on the Internet found that “in 30-80% of the cases,
there is an alternate path with significantly superior quality”
to the default path, where quality is measured in terms of
metrics such as round-trip-time, loss rate, and bandwidth.
This paper outlines the benefits, implementation questions, and open problems associated with using path diversity for streaming media delivery. We begin in Section 2
by describing two basic components of the path diversity
systems we consider: media streaming and media coding.
In Section 3 we survey the benefits of path diversity in a
streaming-media context. Many of these benefits accrue
only for media streaming applications and have no direct
analogy, for example, to benefits associated with classical
wireless diversity. In Section 4 we describe various architectures that support path diversity, including source routing, overlay networks, and IEEE 802.11 wireless infrastructure. Lastly, Section 5 considers a variety of research problems in the design, analysis, and operation of path diversity
systems.
2. BUILDING BLOCKS FOR MEDIA STREAMING
This section describes the general problem of media streaming, focusing in particular on how it differs from transporting generic data. We then provide an overview of conventional single description, scalable, and multiple-description
compression techniques, which are important ingredients in
streaming methods that exploit multiple transport paths.
2.1. Media Streaming
The problem of streaming media such as voice and video
over best-effort packet networks is complicated by a number
of factors. Unlike static content delivery or file download,
streaming involves the simultaneous delivery and playback
of media and is characterized by delay constraints on each
transmitted packet. Packets that arrive after their decoding or display deadline are useless. These inevitable packet
losses are not necessarily fatal; most media streams have
some tolerance to packet loss, albeit limited by the use of
temporally predictive compression. Delay constraints and
tolerance to loss are primary factors that distinguish media
streaming from ordinary data transport.
If the delay constraints are loose relative to typical transport and decoding delays, for example more than a few seconds, then media can be delivered using mechanisms that
resemble file download. Download is convenient for delivering pre-encoded objects of small size such as images or
short audio clips.
Larger objects, such as video clips, require long download times and significant storage at the receiver. A better
approach for delivering these objects is streaming, which involves splitting the object into parts, transmitting the parts
in succession, and decoding and playing back the parts as
they are received.
Streaming is the only option for live (meaning not prerecorded) media objects. Conversational or interactive applications have particularly tight delay constraints, typically
100 milliseconds or less. Meeting these constraints in the
presence of packet losses, queuing delays, and network outages is quite challenging, and provides motivation for some
of the path diversity techniques which follow.
Note that conventional approaches for overcoming packet
loss and network outage in data delivery, such as reliable
transport via TCP, generally are not possible for streaming
because the persistent retransmissions lead to excessive delay. Moreover, many applications (e.g., multicast or broadcast) lack a back channel or other means for retransmissions.
2.2. Media Coding: Single Description, Scalable, and
Multiple Description
Conventional single-description (SD) video coding algorithms,
such as MPEG-1/2/4 and H.261/3/4, compress video into a
single bitstream. In scalable (also called layered) coding,
video is coded into multiple bitstreams, beginning with a
base layer that provides low but usable quality, and one or
more enhancement layers that improve quality. The scalable
coding bitstreams are partially ordered, and different subsets can represent the video at different spatial, temporal, or
amplitude resolutions and fidelities. Applications and networks that support prioritization can exploit scalable coding by assigning higher delivery priorities to coding layers
that are more important. In best-effort networks like the Internet, however, all packets are equally likely to be lost or
delayed. This fundamental mismatch—prioritized data on
a non-prioritized network—makes scalable coding by itself
difficult to exploit using the Internet.
Multiple description (MD) coding is an alternative to
scalable coding that produces two or more bitstreams, or
descriptions, each of roughly equal importance. MD coding has two important properties. First, each description can
be independently decoded to give a usable reproduction of
the original signal. Second, the multiple descriptions contain complementary information so that the quality of the
decoded signal improves with the number of descriptions
received. Note that the first property contrasts with scalable
coding, which has a base layer that is critically important
and if lost renders the other bitstream(s) useless. MD coding enables a useful reproduction of the signal when any
description is received. For example, a simple MD video
coder can be achieved by splitting a video into even and
odd frames and coding them separately.
Repairable MD coding (also known as state recovery
in MD coding) is an extension of conventional MD coding
beneficial for temporally predictively coded media such as
video [1]. Predictive coding causes errors in one frame, due
for example to lost packets, to propagate to potentially many
subsequent frames. In repairable MD coding the decoder
stops error propagation in one description by repairing it using undamaged frames from the other. Such error-resilient
decoding is effective so long as packets are not simultaneously lost for both descriptions — thus coupling nicely with
path diversity. This comparatively recent decoder-side optimization is complementary to the ongoing efforts to improve the rate vs. quality tradeoff of MD encoders, which
typically do not consider partial losses of both descriptions.
The relative compression efficiencies of SD, scalable,
and MD coding depend on the media type. For image and
audio sources scalable coding can be as efficient as SD coding, and many new coding standards, e.g., JPEG-2000, are
scalable. In addition, scalable image or audio coding may
be combined with unequal error protection (UEP) to provide an efficient MD-like system: more redundancy is allocated to the high-priority data relative to the low-priority
data, improving the probability that the high-priority data is
correctly received. For video sources both scalable coding
and MD coding are currently less efficient than SD coding,
however they provide valuable properties for streaming.
3. BENEFITS OF PATH DIVERSITY
Streaming over multiple paths can provide several types of
benefits depending on how the paths are used. In this section
we briefly describe some of these benefits, as itemized in Table 1. For simplicity we assume ideal path diversity where
each path is independent with identical attributes. The case
of realistic paths is examined in Section 5.
Benefits of Path Diversity
Improved (aggregate) BW
Merging
Load Balancing
Splitting
Reduced variability (losses)
Average
Reduced burst losses
Subsampling
Improved error resilience
OR
Reduced outages
AND
Reduced latency (queue diversity)
Min
Table 1. Benefits of path diversity and the associated mathematical operations that intuitively expresses why the benefit occurs.
A straightforward benefit of multiple paths is increased
bandwidth available by using all paths at once. A complementary benefit is load balancing, i.e., decreasing per-path
bandwidth by splitting a stream across multiple paths.
Another benefit of path diversity is reduced variability
of packet losses, e.g., reduced excursions between periods
of no loss and high loss that are common on the Internet.
The end-to-end application sees the average network behavior across the paths, which generally has reduced variability. This reduced variability enables methods for combating
losses, such as forward error correction (FEC), to become
more effective [7, 13].
FEC adds specialized inter-packet redundancy that enables data recovery up to a loss threshold. The amount of
redundancy can be adjusted dynamically to compensate for
changes in the loss rate, but adaptation is problematic when
the network changes quickly: the FEC is inevitably overdesigned (and therefore inefficient) or under-designed (and
therefore ineffective). Combining FEC with long interleaving also helps combat loss variability, but the added delay
makes interleaving unsatisfactory for media streaming. Path
diversity may be viewed as a form of interleaving across
paths rather than across time. Combined with FEC, path
diversity provides benefits similar to time interleaving with
smaller associated delay.
Path diversity also reduces the frequency and length of
burst losses, i.e., losses of consecutive packets. Distributing packets across multiple paths increases the inter-packet
spacing on each path, and therefore for a network congestion event of a given duration fewer packets are lost. Reducing burst losses provides a number of benefits for media
streaming. For example, for video it is easier to recover
from multiple isolated losses than from an equal number of
consecutive losses [1]. For two paths with equal average
packet loss rates, sending even packets on one path and odd
packets on the other has no effect on the end-to-end loss rate
but does reduce burst losses. This provides 0.5–1.0 dB gain
in video quality in certain SD-coded situations [5].
Improved error resilience is a powerful, synergistic benefit that accrues from combining path diversity with repairable
MD coding. To better understand this combination, observe
that repairable MD coding is useful without path diversity
because it reduces error propagation. Conversely, path diversity is useful with non-repairable MD coding because
the distortion remains manageable even if one description
is lost. But both of these “partial” approaches have weaknesses: in the former case burst losses on the single path
simultaneously damage frames in both descriptions and defeat the error-resilient decoder; in the latter, even isolated
errors on the two paths ruin both descriptions. The combination of repairable MD coding and path diversity allows
the system to survive multiple losses—even burst losses—
on the two paths, so long as the losses are not simultaneous.
Path diversity can dramatically decrease the probability
of outage, where an outage is an event where communication along a network path is lost for a sizable length of time.
A path diversity system experiences complete outage only
when all paths are undergoing outages. A system employing MD video coding benefits from such outage protection
because the media stream remains useful even if the traffic
on one path is lost (e.g., 50% packet loss).
Packets sent over independent paths are unlikely to suffer simultaneously from large packet delays. In heuristic
terms, the probability of a bad delay event improves from
to ¾ when the minimum of two path delays governs whether
information is late. We term this effect queue diversity because network delays are often due to backlogged queues,
and to emphasize the benefit of multiple parallel queues.
Queue diversity offers little benefit for loss-intolerant data
because the receiver must wait for all packets from all paths.
In contrast, loss-tolerant delay-intolerant voice or video stream
remain useful when only some packets arrive promptly. The
affected stream may realize the minimum of two or more
path delays (effectively chopping off the long tails in the
end-to-end delay distribution), albeit at some loss in quality.
Examples of such delay reduction include MD voice over IP
(VoIP) [2] and SD video over 802.11 wireless networks [5].
4. ARCHITECTURES FOR PATH DIVERSITY
This section examines different architectures for media streaming with path diversity over packet networks. Specifically,
we discuss: the different media coding approaches and how
they can be coupled with path diversity; two approaches for
path diversity from a single sender to a single receiver; the
use of the distributed infrastructure of a CDN to stream from
multiple senders to a single receiver; the use of multiple
802.11 access points for streaming to a wireless client; and
the use of multiple paths in an ad-hoc wireless network.
4.1. SD versus Scalable versus MD coding?
Effective media streaming with path diversity requires a careful co-design of the compression and the path diversity components. The coding approaches reviewed in Section 2.2
provide a number of tradeoffs, and there are many open research questions as to when each is preferable. In general,
the answer depends crucially on the context, such as the specific media type, source coder, the required playback delay,
the use of retransmissions or FEC, transmission over besteffort or QoS-enabled network, and whether the transmission is unicast or multicast.
For example, while (prioritized) scalable coding is not
a good match to the (unprioritized) best-effort Internet, the
ability to perform prioritized retransmission or unequal error protection at the end hosts can provide a virtual end-toend channel that can better exploit the scalable coding properties. As a further example, the combination of MD coding and path diversity was first proposed for feedback-free
video applications over lossy packet networks where either
the delay of retransmission or FEC is not acceptable (as in
real-time video communications) or where retransmissions
are not possible (as in broadcast or multicast) [1].
4.2. Single-Sender Path Diversity
Path diversity can be achieved over a packet network by
explicitly sending packets through different network paths.
In certain circumstances it is possible to specify the set of
nodes or “source route” for each packet to traverse. Path
diversity can then be achieved by explicitly specifying different source routes for different subsets of packets. While
IP source routing is theoretically straightforward, there are
a number of problems that limit its use in the current Internet, though it may be useful within a private network. It
has been argued that IPv6 may allow source routing to be
adopted more widely.
Another approach to explicitly send different streams
over different paths is to send each stream to a different relay placed at a strategic node in the network, where each
relay performs a simple forwarding operation [1, 13]. The
relay infrastructure appears promising for today’s Internet,
and corresponds to an application-specific overlay network
on top of the conventional Internet. The proposed system
routes traffic through semi-intelligent nodes at strategic locations, thereby providing a service of improved reliability
while leveraging the infrastructure of the Internet. These
paths should be selected to avoid joint bottleneck links, as
discussed in Section 5.
4.3. Multiple-Sender Path Diversity
Content delivery networks (CDNs) have been widely used
to provide low latency, scalability, fault tolerance, and load
balancing for the delivery of web content and more recently
streaming media. CDNs improve end-user performance by
caching popular content on edge servers located closer to
users. This helps prevent server overload and improves scalability. Furthermore, delivering content from a nearby edge
server reduces the request response time, the probability of
packet loss, and total network resource usage.
Another advantage of CDNs is their distributed infrastructure, which can be used to explicitly achieve path diversity between each client and multiple nearby edge servers,
as in Figure 2 [8]. Furthermore, appropriately coupling
MD coding with a CDN can provide improved reliability to
packet losses, link outages, and server failures. This system
is referred to as a Multiple Description Streaming Media
Content Delivery Network or an MD-CDN for short.
An MD-CDN operates in the following manner [8]. The
media is coded into multiple complementary descriptions
which are distributed across different servers in the CDN.
When a client requests a stream, complementary descriptions are simultaneously sent from different nearby servers
through different network paths. This architecture simultaneously reaps the benefits of path diversity and of CDNs.
The multiple servers in a CDN can also provide path
diversity with SD coding and FEC [7]. In this case, the
Received Signal Strength
−20
Power (dBm)
content is replicated on multiple mirror servers and multiple
servers stream disjoint sets of SD and FEC packets to the
single client. As discussed earlier, the reduced variability of
losses afforded by path diversity makes FEC more efficient.
−40
AP1
AP2
−60
−80
0
5000
10000
15000
20000
25000
10000
15000
20000
25000
15000
20000
25000
Average Packet Loss Rate (5 sec window)
20
10
0
0
5000
10
Count
Wireless networks are quite challenging for streaming since
they generally offer time-varying and unpredictable behavior caused by multiple users, interference, propagation effects, and mobility. At the network level, these effects appear as variable delays, losses, and bandwidth. Wireless
links, rather than the wired infrastructure, are often the bottlenecks. As mentioned in the introduction, various forms of
diversity have been used for many years to overcome these
problems, particularly in the cellular environment.
Wireless LANs such as IEEE 802.11 are becoming widespread as they provide simple connectivity and data delivery. One approach to reliable streaming over 802.11 connections, in particular for low-latency or interactive communication, is to exploit the potential path diversity between
each mobile client and multiple 802.11 access points (APs)
in the infrastructure [5]. A mobile 802.11 wireless client is
often in range of multiple APs, each offering a different relationship to the client with respect to distance, obstructions,
multipath, signal strength, contention, available bandwidth,
neighboring interferers, and potential hidden nodes. Figure 3 illustrates the variability seen by a mobile client in an
802.11 network with two access points.
By using multiple paths simultaneously or by switching between multiple paths (site selection) as a function of
channel characteristics, our results indicate that significant
benefits may be achieved by using a wireless path diversity
system compared to a conventional single access point system. These benefits may be achieved using a single client
radio on a single channel [5]; there is no need to perform
physical layer combining or multichannel operation.
Path diversity is also beneficial in ad-hoc wireless networks, where the topology may change or nodes may come
or go. Many ad-hoc routing protocols identify multiple paths
between the sender and receiver, and the use of multiple
paths can improve the reliability of the connection between
the two wireless hosts. The effectiveness of path diversity
over an ad-hoc wireless network for MD image communication is examined in [3].
Average PSNR (dB)
4.4. Path Diversity over Wireless Networks
Percent
30
Number of Burst Losses (5 sec window)
5
0
0
5000
10000
Video Quality
35
30
Oracle
AP1
AP2
25
20
0
10
20
30
40
50
60
70
80
90
Sequence Number
Fig. 3. Performance statistics for path diversity between a
mobile 802.11 wireless client and two APs in the distributed
infrastructure [5]: received packet signal strength variation
(top), average packet loss rate (second), and number of loss
¾ (third) as a function of packet
events of burst length
sequence number for 15-minute packet trace. Bottom plot
illustrates video quality when using only AP1 (red) or only
AP2 (green), and the upper bound on achievable performance by selecting the best AP for every packet (blue).
in effectively designing and using path diversity in streaming media systems.
5.1. Joint and Disjoint Paths
Multiple paths are not guaranteed in practice to be independent. A path may be expressed abstractly as a sequence of
links connected through nodes. Two paths often share links
and nodes. The benefits of path diversity do not depend on
whether paths are completely disjoint, but rather on whether
bottlenecks occur on shared or disjoint portions. Shared
bottlenecks reduce the impact of path diversity, while disjoint bottlenecks do not. Identifying bottlenecks and avoiding them if possible is an important element in effective use
of path diversity. Joint bottleneck detection is currently an
active area of research, e.g., [14].
5.2. How Many Paths to Use?
5. DESIGN, ANALYSIS, AND OPERATION OF
MULTIPATH STREAMING SYSTEMS
The prior sections discuss the potential benefits of path diversity for media streaming and highlight scenarios where
path diversity arises. This section discusses issues that arise
How do diversity benefits scale with the number of paths
and when is it worthwhile to use more than one path? This
complicated question depends on the specifics of the application, the path diversity benefits that one is trying to exploit, and the characteristics of the available paths. For example, for repairable MD video coding, the improved er-
ror recovery arises with two paths, and while increasing
the number of paths helps there are no further jumps in
performance. End-to-end network characteristics improve
in different ways with the number of paths. Total aggregated bandwidth (in principle) increases additively with the
number of paths. Probability of outage decreases exponentially with the number of independent paths. Delay variability,
measured in terms of standard deviation, decreases
as ½ where is the number of independent paths. For
media streaming in a real world scenario, we would expect
that typically a small number of paths would provide a good
balance between complexity and performance.
5.3. Streaming and Packet Scheduling across Asymmetric Paths
Different paths may offer different bandwidths, loss rates,
and delay characteristics. A path diversity system should
compensate for and exploit these asymmetries. The first
step is to estimate the characteristics of each path. Path measurement is an active area of research, and many techniques
developed for single paths may be applied to the multiple
path case, hence we do not cover this topic in further depth.
When paths have different, potentially time-varying, available bandwidths, it is important to adapt the rate across
paths. In the point-to-point case rate adaptation can be performed in a centralized manner. However in the multipointto-point case a distributed algorithm is generally required.
The distributed rate allocation should not only perform the
appropriate adaptation, but also for SD coding should ensure that the different senders send disjoint sets of packets
(no duplication) to the receiver [7].
Paths with different packet loss rates or delays provide
the opportunity to perform packet scheduling across the paths
[9]. For example, more important packets may be sent over
the path with the lower packet loss rate. Packets with tight
delay constraints may be sent over the path with shorter delay. Furthermore, real-time video encoding can be adapted
to react to the time-varying path characteristics and losses [4].
5.4. Selecting the Best Paths or Best Servers
As discussed above, the effectiveness of a pair of paths depends less on their lengths than on their shared vs. disjoint
topology. Of still greater importance is whether bottlenecks
occur on shared portions. Thus, it may be preferable to have
longer paths or longer shared portions of paths if the resulting bottlenecks are not shared. This observation hints at the
complexity of the important basic question: How to select
the best paths to use?
Determining the best paths is important for point-topoint multipath streaming, however its impact becomes even
more apparent when streaming from multiple servers, as in
the case of a CDN. The fundamental problems that arise
when designing and operating a CDN are changed in important ways when using multipath streaming [8]. Three
key CDN problems are:
1. Where to deploy the servers?
2. How to distribute the content?
3. How to select the best server for each client?
In a conventional CDN, where each client receives a stream
from a single server over a single path, the above problems
are solved by minimizing some notion of distance between
a client and a server. However, streaming from multiple
servers to a single client, as in an SD- or MD-CDN, requires
fundamentally different metrics [8]: minimizing distance
from each client to multiple servers while maximizing path
diversity. Note that minimizing distance to multiple servers
and maximizing diversity usually conflict. Accurately evaluating these metrics is necessary for designing and operating a CDN that uses path diversity. In addition, the server
selection problem must be solved for every client request.
5.4.1. Modeling Path Diversity Performance
The above discussion highlights the importance of accurately modeling and predicting the performance of a path
diversity system. An accurate model is needed for selecting
the best subset of paths from a set of possible paths, selecting the best subset of servers from a set of possible servers,
comparing path diversity scenarios, or simply evaluating the
merits of a path diversity system relative to a conventional
single path system.
Determining an appropriate performance model depends
on the benefits (reviewed in Section 3) one intends to capture and on the particular characteristics of the media and
the network. In the following we examine the problem of
accurately predicting SD and MD video quality over a lossy
packet network as a function of path diversity and loss characteristics. Analytical models were proposed in [15, 8] and
are briefly highlighted here; some attributes of these models
may be useful in other scenarios.
Our performance metric is application-level video quality, measured as mean-square error (MSE) distortion at the
receiver. Different types of packet losses affect conventional
SD and MD video differently. The model must therefore
distinguish between isolated packet losses, burst losses of
various lengths, and for MD whether the loss afflicts one or
both descriptions at the same time. To account for losses on
joint vs. disjoint links, each path is modeled as the concatenation of bursty single links (e.g., 2-state Gilbert model),
where each link is separately identified as shared or disjoint.
Modeling media streaming using path diversity is potentially quite complex, as the sender(s) and receiver may
be connected through a wide range of different topologies.
Sample topologies with one and two senders are illustrated
in the top left and top right of Figure 4, respectively. It can
be shown [15, 8], however, that the important end-to-end
properties of a path diversity network are captured using the
dramatically simplified three-subpath topologies shown in
the bottom of Figure 4, where subpaths 1 and 2 are formed
by the disjoint links along the first and second paths, respectively, and subpath 3 is formed by the joint links along both
paths. To summarize, the loss process for two-path path diversity can be accurately modeled by three subpaths, where
the associated burst loss behavior of each subpath is derived
from the corresponding portion of the original path.
Sender
#1
Sender
Receiver
Sender
Subpath 2
(disjoint)
Subpath 3
(joint)
Sender
#1
Receiver
00
01
D12rec
D1drop+D2noloss
10
11
D12drop
Fig. 5. Combined MD performance model: the four states
identify loss conditions, the 16 transition arcs correspond to
different distortions (only 5 are labeled), and the transition
probabilities are determined by the network loss model.
Receiver
Sender
#2
Subpath 1
(disjoint)
D1noloss+D2noloss D1
noloss+D2rec
Sender
#2
Subpath 1
(disjoint) Subpath 3
(joint)
Receiver
Subpath 2
(disjoint)
Fig. 4. Modeling path diversity: Examples of various
path diversity topologies (top), and simplified path diversity
models which preserve the end-to-end characteristics important for accurately predicting video distortion (bottom).
To form a complete application-level performance model,
the packet loss model defined by the simplified topologies in
Figure 4 must be coupled with a distortion model for SD and
MD media streams. The result in the MD case is the finitestate model shown in Figure 5. The four states of the model
represent at each point in time whether both descriptions
are correctly received (state 00), one is correctly received
and one is afflicted by loss (states 01 and 10), or both are
simultaneously afflicted by loss (state 11). The state transition probabilities are derived from the path diversity loss
model, while the distortions accumulated on each transition
are derived from an MD distortion model (see [15, 8]).
The combined model in Figure 5 may be used to compare SD and MD methods in a variety of scenarios. A basic and important comparison is examining conventional SD
over a single path versus MD with two-path diversity. Detailed analysis and results are given in [15, 8], but to summarize for most video sequences, as we vary the fraction
of losses on the joint and disjoint links (with, e.g., 5% endto-end average packet loss rate and average burst length of
1.25 packets) when the losses are on disjoint links then MD
with path diversity is beneficial, however if the losses are
on joint links then the path diversity gain vanishes and SD
sometimes perform better. See [15, 8] for more details.
Additional recent work includes [9] which investigates
rate-distortion optimized packet transmission schedules across
multiple paths, [6] where fast heuristics are presented for
quickly performing path selection when a large number of
candidate path pairs are available, while accounting for the
loss and on-time arrival probability of each path, and [16]
which provides a recent overview of MD video coding and
its transport over multiple paths.
6. SUMMARY
Media streaming with path diversity has gained significant
interest in the last few years, as it provides valuable benefits for overcoming some of the challenges that afflict besteffort packet networks, including dynamic and unpredictable
available bandwidth, delay, and loss rate. Studies have shown
promise for both the single sender to single receiver case,
and the multiple sender to single receiver case as in a streaming media CDN, as well as for both wired and wireless networks. Path diversity helps us take a step closer to feedbackfree video streaming, which could simplify a range of applications from low-latency communication to multicast or
broadcast streaming. The combination of media streaming
and path diversity provides significant promise, and we look
forward to its continuing evolution and adoption.
7. REFERENCES
[1] J.G. Apostolopoulos, “Reliable video communication
over lossy packet networks using multiple state encoding and path diversity,” VCIP, pp. 392–409, January
2001.
[2] Y.J. Liang, E.G. Steinbach, and B. Girod, “Real-time
voice communication over the internet using packet
path diversity,” Proc. ACM Multimedia, pp. 431–440,
Sept/Oct 2001.
[3] N. Gogate, D. Chung, S.S. Panwar, and Y. Wang,
“Supporting image/video applications in a mobile
multihop radio environment using route diversity and
multiple description coding,” IEEE Trans. CSVT, pp.
777–792, Sept 2002.
[4] Y. Liang, E. Setton, and B. Girod,
“Channeladaptive video streaming using packet path diversity
and rate-distortion optimized reference picture selection,” IEEE Fifth Workshop on Multimedia Signal Processing, pp. 420–423, Dec 2002.
[5] A. Miu, J. Apostolopoulos, W. Tan, and M. Trott,
“Low-latency wireless video over 802.11 networks using path diversity,” IEEE ICME, pp. 441–444, July
2003.
[6] A. Begen, Y. Altunbasak, and O. Ergun, “Fast heuristics for multi-path selection for multiple description
encoded video streaming,” IEEE ICME, pp. 517–520,
July 2003.
[7] T. Nguyen and A. Zakhor, “Distributed video streaming with forward error correction,” Packet Video Workshop, April 2002.
[8] J.G. Apostolopoulos, T. Wong, W. Tan, and S.J. Wee,
“On multiple description streaming with content delivery networks,” IEEE INFOCOM, pp. 1736–45, June
2002.
[9] J. Chakareski and B. Girod, “Rate-distortion optimized packet scheduling and routing for media
streaming with path diversity,” IEEE DCC, pp. 203–
212, April 2003.
[10] N.F. Maxemchuk, Dispersity Routing in Store and
Forward Networks, Ph.D. thesis, University of Pennsylvania, May 1975.
[11] A. Banerjea, “Simulation study of the capacity effects of dispersity routing for fault-tolerant real-time
channels,” Computer Communications Review (ACM
SIGCOMM’96), vol. 26, no. 4, pp. 194–205, October
1996.
[12] S. Savage, A. Collins, E. Hoffman, J. Snell, and T. Anderson, “The end-to-end effects of internet path selection,” ACM SIGCOMM, pp. 289–299, October 1999.
[13] T. Nguyen and A. Zakhor, “Path diversity with forward error correction (pdf) system for packet switched
networks,” IEEE INFOCOM, pp. 663–672, April
2003.
[14] D. Rubenstein, J. Kurose, and D. Towsley, “Detecting
shared congestion of flows via end-to-end measurement,” ACM SIGMETRICS, pp. 145–155, 2000.
[15] J.G. Apostolopoulos, W. Tan, S.J. Wee, and G.W. Wornell, “Modeling path diversity for multiple description
video communication,” ICASSP, pp. 2161–2164, May
2002.
[16] Y. Wang, A. Reibman, and S. Lin, “Multiple description coding for video communications,” Submitted to
Proceedings of the IEEE, 2004.